Skip to content

AB-620 study guide ​

Work through the domains in weight order: Integrate, then Plan, then Manage. Drill each table until every scenario cue points to one answer. When two options both seem to work, the right one is usually the governed, lowest-risk choice.

Know cold vs. look up later ​

Know ColdSearchable If Needed
Domain weights: INTEGRATE 40–45, PLAN 30–35, MANAGE 20–25Exact Settings page paths
Generative vs classic orchestration: what changesFull list of 11 trigger names
Teams/M365 → Authenticate with Microsoft onlyManual-auth field names
Which integration fits which need (knowledge / tool / MCP / computer use / agent)Wizard step labels
MCP = tools + resources, generative onlyMCP transport details
Child vs connected agent criteria; Fabric data agent limitsPreview/GA status of each agent type
Knowledge limits: 25 web/SharePoint (gen), 4 (classic); Dataverse 2×15 classicSemantic index file-size limits
Allow ungrounded responses OFF → uncited answer → FallbackCitation rendering per channel
Evaluation: 8 graders and when to use eachContent-safety level numbers beyond 0 = strictest
ALM: custom solution → env vars → managed → pipelines into Managed EnvironmentsPipeline host setup screens

Integrate and extend agents (40–45%) ​

Which integration? ​

NeedAnswerTrap
Ground in third-party content indexed by Microsoft SearchCopilot connector knowledge (user's Entra identity)It's indexed, so it isn't "live"
Answer from live LOB data, no indexPower Platform connector as real-time knowledgeCopilot connector = index
Write/update a recordConnector tool (or REST/MCP tool)Knowledge never writes
Server publishing tools that change over timeMCP (tools + resources; needs generative)Prompts aren't supported
API with an OpenAPI specREST API tool (v2 JSON; v3 auto-converted)Always lands in a solution
No API, only a UIComputer use (5 credits/step; 15 premium)Default = maker credentials
Another agent on another frameworkA2A (endpoint URL; card at /.well-known/agent.json)Not the card URL
Foundry agentProject endpoint + Agent ID; Activity endpoint must be enabled (else 400)Old-portal agents → 404
Fabric data agentAdd it as an agentNo redirect, no / reference, no M365 Copilot, no evaluation
Same team, group toolsChild agent—
Other team, own ALM/model, reuseConnected agent (~30–40 choices is the split point)A main agent with connected agents can't itself be connected

Knowledge know-cold ​

  • Web 25 / SharePoint 25 / Dataverse & connectors unlimited in generative mode. Classic: 4 / 4 per node / 2×15 / 2.
  • More than 25 sources → GPT filtering by description; uploaded files don't count.
  • Bing Custom Search, custom data, Azure OpenAI "on your data" → Classic data in a generative answers node.
  • Azure AI Search: integrated vectorization; one index per connection; Entra (service principal / integrated) over keys; citations need a URL field (metadata_storage_path); VNet via private endpoint + Power Platform VNet support.
  • Official source isn't compatible with generative orchestration.

Azure monitoring ​

  • Settings › Advanced › Application Insights connection string. Options: logging, conversation details (user + text), sensitive properties, node execution events.
  • KQL customEvents; exclude tests with designMode == "False". Block the feature with a data policy on the Application Insights in Copilot Studio connector.

Plan and configure agent solutions (30–35%) ​

Orchestration ​

BehaviorGenerative (default)Classic
Topic selectionName + descriptionTrigger phrases
ToolsAgent picks by descriptionOnly from topics
KnowledgeProactiveFallback / node
Multi-intentChains severalOne topic
Missing inputsAgent asksQuestion nodes
Conversational boosting editsIgnoredUsed
Multiple Topics MatchedNot calledUsed
Custom entities as inputsNot supportedn/a

Controls: End all topics cancels the plan; Reset Conversation clears globals, not history (use Clear variable values › conversation history).

Identity and channels ​

  • Teams + M365 → Authenticate with Microsoft only; others block those channels.
  • Need User.AccessToken or a custom site with SSO → Authenticate manually.
  • No auth = anyone with the link. Sharing-based access control needs Entra (not generic OAuth2).
  • Auth changes apply after publish. D365 Customer Service handoff → no Authenticate with Microsoft.

Agent flows ​

  • Tool-callable only with When an agent calls the flow; return values with Respond to the agent.
  • Billing: Copilot Studio capacity. From a topic = 1 classic answer + actions; from the orchestrator = 1 autonomous action + actions; test chat is free.
  • Cloud flow → agent flow: in a solution, one-way.
  • Multistage approvals: AI stages + manual stages + conditions; no ALM, no sharing; "Analysis failed" = contradictory instructions.

Topic nodes ​

  • HTTP node: headers carry auth; default 30 s timeout; Continue on error → status code + body into variables.
  • Generative answers node: Search only selected sources (no fallback); clear Send a message to post-process into an Adaptive Card; node moderation beats agent moderation.
  • Variables: receive vs return; Date/time, duration, choice and custom-entity values can't be passed; global → topic conversion is one-way; secrets = env var of type secret → Key Vault.
  • Entities: regex is case-sensitive ((?i)); smart matching for fuzzy closed lists; One of multiple entities takes only the first match.

Test and manage agents (20–25%) ​

NeedGrader
No single right answerGeneral quality
Same meaning, any wordingCompare meaning (default pass 50)
Exact code/numberExact match
Near-identical wordingText similarity
Right tool/topic usedTool use
Required termsKeyword match (Any/All)
Harm checkContent safety (0 = strictest)
Business rule, or refusal counts as a passCustom (instructions + labels)
  • Missing expected answer → Invalid. User profile simulates a signed-in user. Automate via REST API or connector. Fabric data agents aren't supported.
  • ALM: default → custom solution (set preferred solution); env vars per target; managed into test/prod; pipeline targets must be Managed Environments; delegated deployments for approval + service principal; no cross-tenant; no skipping stages; import = Upgrade.

Habits to unlearn from older courses ​

Older material showsCurrent Copilot Studio
Typing "trigger phrases" into the topic triggerA one–two sentence description of what the topic does and doesn't do
"Copilot Studio uses Azure CLU"Generative orchestrator picks topics, tools, knowledge and agents by description
Escalate shows a messageTransfer conversation node → engagement hub, va_* context variables
Three trigger types11 types plus priority order (activity → message/event/update/invoke → agent chooses/phrases)
GPT-4.1 / Sonnet 4.5 / Opus 4.5 lineupDefault GPT-5.5 Chat; learn the Deep / General / Auto categories
Graph connectorsCopilot connectors

Traps that look right ​

Tempting answerWhy it's wrong
"Add more trigger phrases" to fix topic confusionGenerative mode → rewrite descriptions to be distinct
"Enable Multiple Topics Matched"Not called under generative orchestration
"Edit Conversational boosting" to change knowledge searchIgnored in generative mode
"Reset Conversation" to clear planner memoryClears globals only
"Add a line to instructions requiring an approved model"Instructions don't select models; prompt model settings do
Copilot connector for "live, no indexing"Copilot connectors are indexing
Add the Fabric data agent via RedirectUnsupported
Deploy unmanaged to production via pipelinesPipelines only deploy managed to targets
General quality for refusal testsPenalizes abstention; use Custom
Web channel security for internal-only accessSecures the embed, not user identity; use Entra auth

Personal study notes written from official documentation. Not affiliated with Microsoft or GitHub. Contains no exam questions.